LG Salles

AI, sales, marketing and development

IA

Tailscale SSH setup: limited VPS access for AI agents

November 16, 2026

Video by David Ondrej · Original title: “Tailscale, Clearly Explained (Beginner's Guide)”Subscribe to the channel

Why it matters

Running another AI agent is easy to justify. Maintaining another machine is where the work accumulates. A service stops, a model setting changes, or an API credential needs replacing. Spread those agents across a laptop and several servers, and every small fix starts with finding the right machine and remembering its setup. David Ondrej’s demonstration makes a useful case for connecting those environments through Tailscale and letting a local agent perform remote maintenance.

The question I would settle first is how much authority that maintenance agent needs. Reading a service log requires a narrower permission set than administering an entire server. Tailscale provides connectivity, while the destination still needs an actual service and appropriate authorization. Joining the network does not automatically supply an SSH server or configure the application you want to reach. Connecting devices.

My starting point would be a two-machine pilot with a specific maintenance job. The useful extension to the video is an acceptance plan: demonstrate that authorized access works, prohibited access fails, and the previous configuration can be restored. That gives you something more useful than a successful installation screenshot when deciding whether to expand the setup.

Key points

  • 00:00 The opening introduces a private network spanning local machines and remote servers. For agents, its practical role is to provide a consistent route for maintenance across those environments.
  • 03:04 Access controls establish boundaries within that network. Adding devices becomes more useful when each device has a defined set of destinations it can reach.
  • 07:43 A coding agent takes over installation work. This demonstrates operational delegation, while leaving a separate question: what evidence will show that the resulting configuration is correct?
  • 12:25 The VPS connection moves onto the Tailscale network. This is where the network path and the SSH authentication method need to be considered separately.
  • 12:56 The demonstration begins restricting public access. A fresh private connection should work before the previous administrative entry point is removed.
  • 15:30 Aperture enters the workflow to centralize access to model providers. Its immediate value is reducing the number of agent environments holding upstream API credentials.
  • 20:11 A dedicated test key receives a spending limit. This makes the experiment easier to contain financially and its usage easier to identify.
  • 24:53 The Hermes installation reaches an end-to-end model response. Completing a package installation and completing a real provider request are separate milestones.
  • 25:24 A remote model change illustrates ongoing maintenance. I would exercise that workflow on one test agent and retain its previous configuration before expanding it.
  • 26:57 Provider records provide additional confirmation. Observing the processed request is stronger evidence of model selection than asking the agent to identify itself.

How to apply

Tailscale SSH setup starts with authentication

First choose between conventional SSH carried over Tailscale and the Tailscale SSH feature. Conventional SSH can use a private Tailscale address while retaining the server’s existing authentication, such as authorized SSH keys. Tailscale SSH manages authentication and authorization through network identity and its SSH policy. Simply connecting two devices does not enable that feature. Tailscale SSH documentation.

Start with your laptop and a disposable Linux VPS. Use recognizable names such as ops-laptop and agent-test-vps. Record the agent’s service name, operating-system user, configuration location, and provider-console recovery procedure. Keep customer data out of this first exercise so you can rebuild the environment freely.

Give the maintenance job a permission budget

Choose a repeatable task: inspect the agent service and restart it when necessary. Describe the required access before implementing it:

| Source | Destination | Required capability | |---|---|---| | Operations laptop | Test VPS | Establish an authorized administrative session | | Agent process | Its own service | Read logs and request a controlled restart | | Test VPS | Model gateway | Send authorized inference requests | | Test VPS | Other servers | None for this pilot |

This is a planning matrix, not a ready-to-paste policy. Network rules restrict reachable destinations and ports. Linux accounts and service permissions restrict what a process can do after connecting. An agent with unrestricted administrator privileges still has unrestricted power inside the server it can reach.

The video discusses ACLs. Current documentation recommends grants for new access policies while continuing to support existing ACLs. Review the whole policy for broad permissions: adding a narrow grant does not cancel an existing rule that already allows wider access. Access control.

Delegate a change with observable completion criteria

I would give the operating agent a bounded instruction:

On the test VPS, identify the agent service and record its current configuration. Configure the private access described in the matrix. Preserve the existing session during validation. Report changed files and rules, test a fresh connection, and document how to restore the previous configuration. Limit the work to this server.

After installing Tailscale on both devices, inspect membership with tailscale status. Run tailscale ping agent-test-vps to examine connectivity, then test the actual SSH session separately. A successful Tailscale ping alone does not establish that the application service is reachable. Connections can be direct or relayed, with encryption between devices in either case; network conditions determine which path is available. Connection types.

Once a fresh private administrative session works, adjust public inbound access according to the server’s actual responsibilities. A host serving a public website or webhook needs a different policy from a dedicated private agent machine. Inventory those services before delegating firewall changes.

Put provider credentials behind the model gateway

In the demonstrated arrangement, Aperture receives model requests and adds the upstream provider credential. This lets the agent operate without storing that provider key in its own configuration. Treat it as a model gateway rather than a general store for database passwords, backup credentials, and every other server secret. How Aperture works.

Configure one provider and one model first. Send a short request, find its corresponding record, and inspect the VPS for leftover copies of the original key in configuration files or shell history. A private connection to the gateway also does not make inference local: the request still travels onward to the selected model provider.

Test the boundaries and the way back

A useful pilot should produce evidence for each outcome below:

| Proposed test | Expected outcome | |---|---| | Fresh session through the authorized route | Connection succeeds | | Unauthorized source attempts to use the service | Access is denied | | Short model request | Response matches a recorded request | | Model change on the test agent | Configuration and provider records agree | | Restore the previous configuration | Agent responds again |

Record elapsed time and human interventions across several maintenance runs. That will show whether the setup reduces recurring work enough to justify maintaining the network policy and gateway. I would expand to another server only after those results are repeatable.

Check the applicable commercial terms as part of that decision. Tailscale’s free Personal plan is for non-commercial use, so a small team should not assume it qualifies based on headcount alone. Include networking, VPS hosting, and model usage when comparing the pilot’s cost with the maintenance time it saves. Tailscale plans.

Resources mentioned

  • Tailscale (paid) · Rede privada demonstrada; plano Personal gratuito para uso não comercial e planos pagos para empresas.
  • Aperture by Tailscale (paid) · Gateway de modelos demonstrado; condições variam por plano e modalidade, com consumo de modelos cobrado.
  • Hermes Agent · Agente instalado no VPS; código aberto e gratuito para hospedar, com infraestrutura e modelos à parte.
  • cmux · Terminal usado na demonstração; versão local gratuita para macOS e modalidades pagas adicionais.
  • Codex (paid) · Agente usado para instalar e administrar o ambiente; acesso e limites dependem do plano ou modalidade.
  • OpenRouter (paid) · Provedor configurado no Aperture; cobrança conforme os modelos utilizados, com opções gratuitas no catálogo.
  • Hostinger VPS (paid) · Hospedagem demonstrada e patrocinadora do vídeo; link oficial sem código de afiliado.
  • Tailscale Agent Network — materiais de David Ondrej · Página oficial dos materiais anunciados no vídeo; solicita nome e e-mail.
  • David's Agent Skills · Página oficial das skills recomendadas pelo autor; solicita nome e e-mail.